Bayes estimators for phylogenetic reconstruction Ruriko Yoshida Origins

Bayes estimators for phylogenetic reconstruction Ruriko Yoshida

Origins of Species Figure 19. 1 Genomes 3 (© Garland Science 2007)

Tree (or web) of life eukarya History of life is written in the genes. bacteria Mount DW (2001) Bioinformatics: sequence and genome analysis. Cold Spring Harbor Laboratory Press, Cold Spring Harbor, New York archaea

Goals of molecular phylogenetics Read the evolutionary history of organisms genes Stechmann A, Cavalier ST (2002) Rooting the eukaryote tree by using a derived gene fusion. Science 297: 89 -91

Figure 19. 20 Genomes 3 (© Garland Science 2007)

Why mathematics and statistics? To analyze huge/messy data sets we need to use statistical methods. Mathematical tools to organize / analyze the data sets. We need algorithms to compute them efficiently. Too bad you have to use it!

Gene tree on a Species tree Maddison WP (1997) Gene trees in species trees. Systematic Biology 46: 523 -536

Phylogenetic tree reconstruction The maximum likelihood estimation (MLE) methods: These methods describe evolution in terms of a discrete-state continuous-time Markov process. The Balanced Minimum Evolution (BME) method: This is a distance based method and weighted Least Square method. The Neighbor. Joining (NJ) method is a greedy algorithm of the BME method (Steel and Gasquel, 2008). Bayesian inference for trees: Use Bayes Theorem and MCMC to estimate the posterior distribution rather than obtaining the point estimation.

Closeness Tree uncertainty is a pervasive issue in phylogenetics. To help cope with tree uncertainty, bootstrapping and Bayesian sampling methods provide a collection of possible trees instead of a single tree estimate. Using bootstrapping or Bayesian sampling, one common practice is to identify highly supported tree features (e. g. splits) which occur in almost all the tree samples. Highly supported features are regarded as likely features of the true tree. Similarly, in simulation studies it is common to judge reconstruction methods based on how close they get to the true tree.

Bayes estimators This leads us to ask whether reconstruction accuracy (i. e. closeness to the true tree) can be improved, by attempting to directly optimize accuracy. Even though the true tree is unknown, we can still optimize reconstruction accuracy using a Bayesian approach. In the Bayesian view, the true tree is a random variable T distributed according to the posterior distribution P(T|D), where D is input data such as sequence data.

Bayes estimators If d() measures distance between trees, and T’ is a tree estimate, then the expected distance between T’ and the true tree is ET~P(T |D) d(T, T’). Thus, to maximize reconstruction accuracy, we should choose our tree estimate to be T* = argmin. T’ ET~P(T |D) d(T, T’) where T* is known as a Bayes estimator.

Tree Distances Many popular distances between trees can be easily expressed as a squared euclidean distance, after embedding trees in an appropriately chosen vector space. Important examples include Robinson-Foulds distance (symmetric difference), quartet distance, and the squared path difference. With Robinson--Foulds distance, Holder showed the Bayes estimator with the RF distance is the majority-rule consensus tree (Holder, Systematics Biology 2008). We focus on squared euclidean distances, specifically the squared path difference.

Path Difference Metrics The path difference metric is a topological distance, i. e. it only depends on the topologies of the tree. The path difference metric vp(T) in Rn(n-1)/2, where n is the number of taxa, is the integer vector whose the ijth entry, i < j, counts the number of edges between leaves i and j in the tree T. The path difference metric was studied in (Steel and Penny, 1993). The squared path difference is dp(T', T) = || vp(T) – vp(T') ||2.

Example For the trees T 1 and T 2, we have: vp(T 1) = (2, 3, 4, 4, 3, 3, 2) vp(T 2) = (2, 4, 4, 3, 2, 3, 3) dp(T 1, T 2) = || vp(T 1) – vp(T 2) ||2 = 6 T 1 T 2

Optimization We want to find the optimal solution T* such that T* = argmin_T’ ET~ P(T | D) dp(T, T'). Since the number of tree topologies on n taxa grows exponentially in n, computing the Bayes estimator T* under a general distance function can be computationally hard (NP hard).

Hill Climbing Techniques However, hill climbing techniques such as those used in ML methods often work quite well in practice for tree reconstruction. Hill climbing techniques can similarly be used to find local minima of the empirical expected loss. Hill climbing requires a way to move from one tree topology to another. Three types of combinatorial tree moves are often used for this purpose; Nearest Neighbor Interchange (NNI), Subtree-Prune-and. Regraft (SPR), and Tree-Bisection-Reconnect (TBR). Here we use NNI moves.

Nearest Neighbor Interchange Here we assume unrooted trees (if we want to have a rooted tree we add an outgroup and make it as unrooted tree with n + 1 taxa). A, B, C, D are clades

Nearest Neighbor Interchange So we can think of this as a tree with 4 leaves and since there are only 3 tree topologies we swap this tree with one of the following two different tree topologies.

Simulations For simulated data, we used the first 1000 examples from the data set presented in (Guindon and Gasquel, 2003). Trees on 40 taxa were generated according to a Markov process. For each generated tree, 40 homologous sequences (no indels) of length 500 were generated, under the Kumura two-parameter (K 2 P) model, with a transition/transversion ratio of 2. 0. Specifically the Seq-Gen program was used to generate the sequences. The data is available from the website http: //www. atgc-montpellier. fr/phyml/datasets. php

Simulations For each set of homologous sequences D in the simulated data, we used the software Mr. Bayes to obtain 15, 000 samples from the posterior distribution P(T|D). Specifically, we ran Mr. Bayes under the K 2 P model, discarded the initial 25% of samples as a burn-in, used a 50 generation sample rate, and ran for 1, 000 generations in total. We computed a consensus tree computed via Mr. Bayes, a ML tree estimate for each data set, using the hill climbing software PHYML as described in the paper. We also computed a NJ tree using the software PHYLIP, using pairwise distances computed by PHYLIP.

Siulations

Simulations

Acknowledgments P. Huggins, W. Li, D. Haws, T. Friedrich, J. Liu, and R. Yoshida. Bayes estimators for phylogenetic reconstruction, accepted conditionally in Systematic Biology. Available at http: //arxiv. org/abs/0911. 0645. Collaborators: - Peter Huggins (Fall 2010, UK) - David Haws (Stat, UK) - Wenbin Li (CS, UK) - Thomas Friedrich (Stat, UK) - Jinze Liu (CS, UK) • • NIGMS NSF

“Are you done yet? ” Thank you for your attention!